Ribhu Lahiri 🪢

kernels compiled for your exact chip. gentoo won

LearnHarness: an agent harness for learning things. The goal is strong, usable understanding, with mental effort focused on direct interaction with the material rather than managing the learning process. The best explanations and the best practice problems depend on what you already understand and how you understand it. So the agent needs a detailed picture of that understanding, continually updated as you learn. TEACHING 1. Probe The agent probes your understanding of everything the lesson depends on, building a detailed picture of your mental models and ways of thinking. 2. Plan The agent reasons out the minimal dependency tree from what you already know to what you want to understand. 3. Teach Starting from what you already understand, the agent teaches one reasoning step at a time. The aim is for each step, and eventually the goal, to feel obviously true. At each step: i) You predict the next move. ii) The agent explains it. iii) The agent checks your understanding. Predictions and checks help surface errors and confusions immediately, so they can be resolved before later steps build on them. PRACTICE Following the reasoning and having the right ideas come to mind when needed are different abilities. Practice makes understanding usable by broadening and strengthening cues for recall. Problems adapt to target what needs strengthening and to stay at an ideal difficulty level: hard enough to require real struggle, but not so hard that you can’t make progress. When you get stuck, the agent aims to give the smallest possible hints needed to unblock your thinking without replacing it. SHARED RECORD Teaching sessions and imported notes are distilled into propositions and reasoning chains. This compact, persistent record is shared between teaching and practice. It renders as a graph. APP Subagents research, read PDFs, create diagrams and build interactive demos. Free, open-source desktop app. Use your own keys and models. Learning records are local plain files. System prompts are visible and editable. github.com/amosblomqvist/lea…

Can LLMs improve their own harness? Self-improving harnesses are a hot topic. And one interesting question is whether LLMs can generate and improve them on their own. They find that two full rewrites of the harness raise held-out scores by only 2.80 points, and the 95% interval crosses zero. Each frontier model writes a complete harness from a fixed coding bootstrap, plays five development games through it, and then rewrites its whole codebase twice. Every version is scored on 20 games that never influence the rewrites. Across 27 lineages, first harnesses range from zero to near-ceiling. The revisions add very little. Gains on the development games overstate the held-out gains, and revisions that cut inference cost often lower the score. Human-steered systems still score higher on the same games, so better designs exist. The models just do not find them reliably on their own yet. arxiv.org/abs/2610.30499 Chat with Paper: academy.dair.ai/papers/can-l…

Dang, we saw it coming, but dang > Arc is sunsetting on December 22, 2026. arc.net/

i honestly thought this was satire

Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

on that note, I'll be hosting Hack for Humanity on 10th October. see ya folks - luma.com/h4h-delhi

Here’s to the wanting. Here’s to the dreaming. Here’s to everyone.